TRUST: an large language model-based dialogue system for trauma understanding and structured assessments

Area of Science:

• Artificial Intelligence in Healthcare
• Clinical Psychology
• Natural Language Processing

Background:

• Large language models (LLMs) are increasingly used in healthcare, but not for diagnostic interviews.
• Mental healthcare accessibility remains a challenge.
• Standardized diagnostic interviews are crucial for accurate assessment.

Purpose of the Study:

• To develop an LLM-powered dialogue system (TRUST) that replicates clinician behavior for diagnostic interviews.
• To bridge the gap in mental healthcare accessibility.
• To conduct formal diagnostic assessments for post-traumatic stress disorder (PTSD).

Main Methods:

• Introduced TRUST, a framework of cooperative LLM modules.
• Utilized a Dialogue Acts schema for clinical interview response generation.
• Developed a patient simulation approach using real-life interview transcripts for efficient testing.

Main Results:

• Designed comprehensive evaluation metrics for agent and patient simulation perspectives.
• Expert evaluations confirmed TRUST performs comparably to real-life clinical interviews.
• Demonstrated clinical quality approaching that of human clinicians.

Conclusions:

• The TRUST framework shows potential to significantly enhance mental healthcare availability.
• The system's performance approaches clinical quality, with opportunities for future communication improvements.
• LLM-powered dialogue systems can effectively support diagnostic processes.